OpenAI has expanded its GPT-6 family with two faster, lower-cost models aimed at everyday professional work and agents.
GPT-6 Sol targets demanding coding, computer-use, and multi-step workflows. GPT-6 Luna is optimized for focused, high-volume tasks such as extraction, summarization, classification, and lightweight agent steps.
Both models are available through the API, Codex, and ChatGPT Work. GitHub has also added them to Copilot across major IDEs and its cloud coding agent.
The price shift matters
For standard API requests with no more than 272,000 input tokens:
- GPT-6 Sol: $2 input and $10 output per million tokens;
- GPT-6 Luna: $0.10 input and $0.50 output per million tokens; and
- cached input costs $0.20 for Sol and $0.01 for Luna.
That represents a 50% reduction from GPT-5.6’s promotional pricing. OpenAI has also improved prompt caching for long conversations and agents, including the ability to change reasoning effort or available tools without invalidating earlier cached context. Requests above the 272,000-token threshold use higher rates.
OpenAI reports that Sol makes roughly half as many factual mistakes as GPT-5.6 Sol on its internal evaluation. It scored 68.8% on DeepSWE at maximum reasoning effort, while Luna scored 66.6%. These are vendor-reported results, but they reinforce the central product story: capable agent models are becoming substantially cheaper to run.
Why it matters for builders
Long-running agents repeatedly process instructions, repository context, tool definitions, intermediate results, and conversation history. Model cost therefore compounds much faster than it does in a simple chatbot.
GPT-6’s pricing and caching improvements make several architectures more practical:
- route complex planning and validation to Sol;
- use Luna for focused subtasks and high-volume tool loops;
- preserve cached context while adjusting reasoning effort;
- run more evaluations and retries without immediately exhausting the budget; and
- reserve GPT-6 Astra for the comparatively small number of tasks requiring maximum capability.
The important shift is not simply that two new models are available. Developers can now design agents around a genuine capability-and-cost ladder instead of assigning every step to the same expensive frontier model.